Re: -O2 compiler bug on amd64? (Re: new TCL vs. current-6)

From: Michael Johnson <ahze_at_ahze.net>
Date: Tue, 14 Dec 2004 13:33:53 -0500
On Dec 14, 2004, at 1:25 PM, Jeremy Messenger wrote:

> On Tue, 14 Dec 2004 14:24:53 +0100, Anthony Ginepro 
> <anthony.ginepro_at_laposte.net> wrote:
>
>> Le Mardi 14 décembre 2004 à 13:39 +0100, Alexander Leidinger a écrit :
>>> Zitat von Kris Kennaway <kris_at_obsecurity.org>:
>>>
>>> > > Peter speculated a problem when the world is built with -O2 
>>> (with or
>>> > > without -fno-strict-aliasing).  I haven't been able to confirm 
>>> this by
>>> > > testing with a world built with -O, because I only have one 
>>> working
>>> > > amd64 package builder.
>>> >
>>> > Looks like building the port against an -O world has fixed the
>>> > problems.  This suggests a compiler bug with -O2 on amd64.
>>>
>>> I "hope" this is the case, not only on amd64. Background: I see 
>>> coredumps
>>> of gst-register on my system (x86). I've recompiled all ports, with 
>>> and
>>> without -Os, with and without -fno-strict-aliasing, but it still
>>> segfaults. I also recompiled world with "-Os -fno-strict-aliasing" 
>>> instead
>>> of my usual "-Os". I haven't tested plain "-O" for the world yet, 
>>> but I
>>> think the problem is somewhere in the system libs...
>>>
>>> Bye,
>>> Alexander.
>>
>> Some more information :
>> gst-register coredumps if gstreamer-plugins is compiled with 
>> CPUTYPE=p4.
>> eclipse also coredumps if gtk2 is compiled with CPUTYPE=p4.
>> Both ports don't depend on optimization level (-O2 or -O).
>
> gstreamer-plugins is weird.
>
> Crash:
> =======================
> CPUTYPE?=athlon-xp
> CFLAGS= -O -pipe
> =======================
>
> No crash:
> =======================
> #CPUTYPE?=athlon-xp
> CFLAGS= -O -pipe
> =======================
>
> No crash:
> =======================
> CPUTYPE?=athlon-xp
> CFLAGS= -O2 -pipe
> =======================
>
> No crash:
> =======================
> Add '-fgcse' in gstreamer-plugins and it will fix everything for 
> athlon-xp, but not p4. The '-fgcse' is automatic add in when you use 
> -O2, so it's why it doesn't crash with -O2.
> =======================
>
> I personal have no idea if it's bug in FreeBSD or gstreamer-plugins, 
> but there is no issue if it's in Linux.
>
> On another note, there has different thread in freebsd-threads about 
> 'SSE vs. stack alignment vs. pthread'. It included a libc patch that 
> fix for an app for SSE and 16-byte aligned stuff on p4.
>
> Cheers,
> Mezz

The crash is also only from one plugin from gstreamer-plugins, modplug. 
And modplug isn't *really* needed. Why don't we just get rid of modplug 
and and a optional knob for it?
This will solve everyone's problem with it.

Michael
>
>> Anthony.
>
>
> -- 
> mezz7_at_cox.net  -  mezz_at_FreeBSD.org
> FreeBSD GNOME Team
> http://www.FreeBSD.org/gnome/  -  gnome_at_FreeBSD.org
> _______________________________________________
> freebsd-amd64_at_freebsd.org mailing list
> http://lists.freebsd.org/mailman/listinfo/freebsd-amd64
> To unsubscribe, send any mail to 
> "freebsd-amd64-unsubscribe_at_freebsd.org"

Received on Tue Dec 14 2004 - 17:34:08 UTC

This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:38:24 UTC